There was a discussion on this benchmark https://getomni.ai/ocr-benchmark couple of weeks ago here > https://news.ycombinator.com/item?id=43118514
replies(1):
To give an example, some models like mixtral-ocr only try to convert a document to markdown format. You have to use another LLM on top of it to get the final result. Some VLM’s directly give structured information like key fields from documents like invoices, but you have to either add business rules on top of it or use some LLM as a judge kind of system to get sense of which output needs to be manually reviewed or can be taken as correct output. No benchmark attempts to measure the actual rate of automation you can achieve.
We have tried to solve this problem with a benchmark that is only applicable for documents/usecases where you are looking for automation and its trying to measure that end to end automation level of different models or systems.
We have collected a dataset that represents documents like invoices etc which are applicable in processes where automation is needed vs are more copilot in nature where you would need to chat with document. Also have annotated these documents and published the dataset and repo so it can be extended.
Here is writeup: https://nanonets.com/automation-benchmark Dataset: https://huggingface.co/datasets/nanonets/nn-auto-bench-ds Github: https://github.com/NanoNets/nn-auto-bench
Looking for suggestions on how this benchmark can be improved further.
One question on the "Automation" score in the results, is this a function of extraction accuracy vs the accuracy of the LLM's "confidence score". I noticed the "accuracy" column was very tightly grouped (between 79 & 84%) but the automation score was way more variable.
And side note: is there an open source Mistral benchmark for their latest OCR model? I know they claimed it was 95% accurate, but it looks that was based on an internal evaluation.
Good way to think about automation is recall at high precision which is what you need for true automation where you don't worry about documents that are very likely to have correct results and focus on manually correcting documents likely to have errors.
The reason accuracies are tighly grouped but not the automation is because these models are trained to be accurate but not necessarily predictable, where there is no real way to get confidence score calibrated.
Couldn't find the benchmark mistral used as well
One thing I’ve noticed in many benchmarks, though, is the potential for bias. I’m actually working on a post about this issue, so it’s top of mind for me. For example, in the omni benchmark, the ground truth expected a specific order for heading information—like logo, phone number, and customer details. While this data was all located near the top of the document, the exact ordering felt subjective. Should the model prioritize horizontal or vertical scanning? Since the ground truth was created by the company running the benchmark, their model naturally scored the highest for maintaining the same order as the ground-truth.
However, this approach penalized other LLMs for not adhering to the "correct" order, even though the order itself was arguably arbitrary. This kind of bias can skew results and make it harder to evaluate models fairly. I’d love to see benchmarks that account for subjectivity or allow for multiple valid interpretations of document structure.
Did you run into this when looking at the benchmarks?
On a side note, Doctly.ai leverages multiple LLMs to evaluate documents, and runs a tournament with a judge for each page to get the best data (this is only on the Precision Ultra selection).
Example is if you are only converting lets say an invoice into markdown, you can introduce bias wrt ordering etc. But if the task is to find out invoice number, total amount, number of line items with headers like price, amount, description, in that case you can compare two outputs without a lot of bias. Eg even if columns are interchanged, you will still get the same metric.
We chose this method for all the same reasons you highlight. Text similarity based measurements are very subject to bias, and don't correlate super well with accuracy. I covered the same concepts in the "The case against text-similarity"[1] section of our writeup.
I’m vibe coding a little macOS OCR app since last weekend, and I’m really happy with the results so far. This is my first app, so fingers crossed. If it becomes feature-complete and polished enough, I’m considering open sourcing it. There’s still a long way to go, though.
[1] https://developer.apple.com/documentation/vision/vnrecognize...
If you create a benchmark and your product outperforms everything else, it could mean many things. Overfitting being one of them.
Which is why publishing open benchmarks is first step where there is public scrutiny around whether then benchmark itself irrespective of the results is fair. In the end, the end user will choose the benchmark that's best fit for their usecase or mostly will create a variation of their own, do their own unbiased evaluations.
There are also:
* swiftocr - https://github.com/fny/swiftocr
* macos-vision-ocr - https://github.com/bytefer/macos-vision-ocr
[1]: https://docs.nanonets.com/docs/consumption-based-billing#exa...